vocabulary

name	vocabulary
description	Use when writing, reviewing, or naming any responsibility-laundering book artifact — pick the canonical noun or verb from this registry rather than coining a synonym. Loaded by NLPM's scorer and checker when R51 is enabled in .claude/nlpm.local.md.
version	0.1.0

Responsibility Laundering Book — Domain Vocabulary

The canonical noun-and-verb set for this book project. Every artifact name, prose description, and rule wording should draw from this registry. Synonyms are flagged by /nlpm:check and penalized by /nlpm:score when R51 is enabled.

Status: seeded by /nlpm:vocab-init on 2026-05-25. The seed combines (a) literary warrant from analysis/vocabulary-extract/ and (b) the book's domain taxonomy already codified in .claude/rules/00-five-values.md through .claude/rules/04-style-guide.md and the AGENTS.md core diagnostic.

Adopter must prune, define deprecation pairs, and confirm scopes before R51 enforcement produces useful signal. Re-run analysis/scripts/extract-vocabulary.py after artifact changes to refresh.

---

Scopes (P1 of vocabulary-design-principles.md)

This project has a single internal scope. There is no auditor scope because no .github/workflows/auditor-*.yml files exist.

Scope	Paths	Description
internal	.claude/commands/, .claude/agents/, .claude/skills/, .claude/docs/, book/, AGENTS.md	What this project does to its own artifacts: build case files, draft chapter briefs, draft prose, audit evidence, red-team claims, handoff between crew roles.

Verbs (literary warrant from corpus + book methodology)

The corpus is prose-heavy, so the extractor surfaced few action verbs from filenames alone. The table below combines extractor output (frequency column) with the verbs the book methodology actually uses (frequency 0 marks book-methodology verbs not yet in filename warrant).

internal scope

Canonical verb	Frequency (filename + description)	Deprecated synonyms (fill in)
build	4	make, produce, create (only when result is an artifact)
audit	2	check, review (only when output is a usable/weak/unusable verdict)
draft	5	compose, pen, author, write (now deprecated; was the dominant negative-constraint form)
revise	0	rewrite, polish, edit (only when input is a prior draft)
compile	0	assemble, gather
verify	0	confirm, double-check
red-team	0	attack, stress-test, critique
fact-check	0	verify (only when subject is a single claim)
grade	21	rank, score, rate, mark (only when output is an A/B/C/D label; mark moved here from flag per drift report)
flag	0	note, warn (mark moved to grade per drift report)
gate	0	block, hold, stop (only when output is a draft-readiness decision)

Nouns (literary warrant from corpus + book taxonomy)

Artifact-class nouns

Canonical	Frequency	Definition
case file	warrant from commands:filename, skills:dirname	Structured record per .claude/skills/case-file-method/SKILL.md — eight-question diagnostic, evidence grade, source ledger, handoff.
chapter brief	warrant from commands:filename	Plan-of-prose per .claude/skills/chapter-blueprint/SKILL.md — scenes, thesis, case hierarchy, counterargument, evidence gaps.
chapter draft	book methodology	Prose output from Wayne that turns a chapter brief into the chapter rhythm in .claude/rules/04-style-guide.md.
source ledger	warrant from skills:dirname	Per-claim citation log per .claude/skills/source-ledger-discipline/SKILL.md with provenance and interest disclosure.
responsibility chain	book methodology	The control → benefit → knowledge → preventability → record control → cost-bearing map per .claude/skills/responsibility-chain-mapping/SKILL.md.
handoff	book methodology (AGENTS.md "Default artifact standard")	The Owner / Purpose / Evidence grade / Assumptions / Open questions / Handoff footer every substantial output ends with.
red-team memo	book methodology	Output of the /red-team command — strongest objection, likely critic attack, correction path, safer wording.
proposal pack	warrant from commands:filename	Output of /proposal-pack — pitch, titles, overview, audience, comps, chapter summaries, sample-chapter strategy.
claim	book methodology (.claude/rules/02-evidence-grades.md)	A single assertion that takes an A/B/C/D evidence grade.
scene	book methodology (.claude/rules/04-style-guide.md)

Case-taxonomy nouns (load-bearing; do not paraphrase)

Per .claude/rules/01-case-taxonomy.md these four labels are non-substitutable and must be used verbatim:

Canonical	Forbidden substitutes
pure scapegoat	innocent victim, wrongly accused
partial scapegoat	mostly guilty, lower-level fall guy
system/object alibi	procedural cover, system blame, process scapegoat
cost-bearing goat	collateral damage, silent victim

Role-nouns (agent files; paired with their dominant verbs)

Role-noun	Frequency (agents:filename)	Paired verb(s)
reviewer	1	review (Alan, parameterized across 6 domains)
researcher	4	build (source packets), cite (Shirley/Selina/Warren/Loki)
director	2	assign, gate (Delon for research, Stephen for fact-check)
strategist	1	position, test (Blair, combined proposal + audience)
editor	1	red-team (Laura)
architect	1	place, sequence (Bonnie)
lead	1	draft, revise (Wayne)
chief	1	assign, consolidate, decide (Jerry)
counsel	1	flag, gate (Nancy — legal risk)

Diagnostic-question nouns (from AGENTS.md core diagnostic)

Every case file answers in this exact wording:

Canonical	Forbidden substitutes
control	power over, decision authority
benefit	gain, upside
knowledge	awareness, what they knew
preventability	avoidability, could have stopped
record control	paper trail control, archive ownership
cost bearer	harmed party, loser

---

How to extend

1. Re-run extraction: python3 ~/.claude/plugins/cache/xiaolai/nlpm/0.8.23/analysis/scripts/extract-vocabulary.py --root . --scopes analysis/vocab-init-scopes.json --out analysis/vocabulary-extract/
2. New terms appear in analysis/vocabulary-extract/summary.md.
3. Add a row to the right table above. Cite at least one file as evidence.
4. To deprecate a synonym, list it in the canonical term's "deprecated synonyms" column and mirror the entry in registry.yaml.

Adopting R51

R51 is off by default. To enable vocabulary drift detection on this project, add to .claude/nlpm.local.md:

rule_overrides:
  R51:
    enabled: true
    vocabulary_skill: .claude/skills/vocabulary/

Without this opt-in, R51 contributes zero penalty regardless of artifact contents.

See also

• NLPM's own vocabulary skill (the template this was seeded from): skills/nlpm/vocabulary/SKILL.md in the NLPM repo.
• The six principles: analysis/vocabulary-design-principles.md in the NLPM repo.
• This project's case taxonomy: .claude/rules/01-case-taxonomy.md.
• This project's evidence grades: .claude/rules/02-evidence-grades.md.
• This project's style guide (verbs, banned phrases, chapter rhythm): .claude/rules/04-style-guide.md.

Context: Wayne is drafting a chapter and types "the algorithm did it" to describe a system/object alibi. output: R51 flags "algorithm did it" as a substitute for the canonical `system/object alibi`. Suggested rewrite: "the algorithm is the named cause; the laundering moves blame from the team who deployed it to the model itself" — pulls the canonical taxonomy term into the prose so a later /source-audit can grade the surrounding claims. Context: A new researcher proposes adding "framework" as an artifact-class noun. output: Refused pending warrant. The term appears 1× in agent filenames (`jade-visual-framework-designer`), but no skill, command, or rule defines it as a distinct artifact class. Added to `rejected_pending_warrant:` in `registry.yaml` with a note pointing to the proposal. Will be reconsidered if frequency rises above 3 with a definition in a rule file.